# FastH3 4-step Preview (FastVideo) — card for AI agents

Four-step distilled MiniMax H3 checkpoint that generates synchronized video and audio from text.

Category: AI video generators and platforms · maker FastVideo (Hao AI Lab) · checked 2026-10-27

## Price
- Price: Free
- Free tier: Open weights, free to download; self-hosted. Needs the MiniMax H3 base model.

## License
MiniMax H3 Community License Agreement — https://huggingface.co/FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree/blob/main/LICENSE

- commercial: **conditions** (Commercial use with conditions)
  > You may not use, reproduce, modify, distribute, or display the MiniMax H3 Works or any of their Outputs or results outside the Applicable Territory.
  (https://huggingface.co/FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree/blob/main/LICENSE)
- attribution: **required** (Credit required)
  > on the user interface of commercial product or service that uses MiniMax H3 or MiniMax H3 Works.
  (https://huggingface.co/FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree/blob/main/LICENSE)
- in app: **conditions** (Ship inside your app with conditions)
  > you must, before making that product or service available and throughout its operation, implement, maintain, test, and periodically review reasonable and proportionate technical and organizational safeguards
  (https://huggingface.co/FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree/blob/main/LICENSE)
- youtube: **conditions** (Monetized YouTube with conditions)
  > MiniMax claims no rights over the Outputs you generate.
  (https://huggingface.co/FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree/blob/main/LICENSE)
- ai training: **banned**
  > You may not use the MiniMax H3 Works or any of their Outputs or results to improve any other artificial intelligence model
  (https://huggingface.co/FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree/blob/main/LICENSE)

## Rules to follow
- Use is allowed only inside the Applicable Territory: it excludes the EU, UK, South Korea and the USA.
- Show "MiniMax H3" in the interface of commercial products; ask MiniMax for a license above 20 million USD yearly revenue.
- Do not use the weights or their outputs to train or improve another AI model.
- Hosting generation for third parties needs safeguards under the Acceptable Use Policy.

## How to connect
- Download only
- Repository: https://github.com/hao-ai-lab/FastVideo
- Website: https://huggingface.co/FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree

## Notes
Preview checkpoint; tested defaults use four B200 GPUs. Revenue above 20 million USD per year needs separate written authorization from MiniMax.

## Numbers (AgentGid data)
- Gid Score: 25
- GitHub stars: 4,829

## Lines to paste into your prompt or AGENTS.md

```
- FastH3 4-step Preview (FastVideo) (AI video generators and platforms). Four-step distilled MiniMax H3 checkpoint that generates synchronized video and audio from text. Price: free; commercial use with conditions; credit required; ship inside your app with conditions; no AI training on its content.
  Connect: download from https://huggingface.co/FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree (no API).
  Rule: Use is allowed only inside the Applicable Territory: it excludes the EU, UK, South Korea and the USA.
  Rule: Show "MiniMax H3" in the interface of commercial products; ask MiniMax for a license above 20 million USD yearly revenue.
  Rule: Do not use the weights or their outputs to train or improve another AI model.
  Rule: Hosting generation for third parties needs safeguards under the Acceptable Use Policy.
  License: https://huggingface.co/FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree/blob/main/LICENSE (checked 2026-10-27; details: https://agentgid.com/assets/fastvideo/)
```

## Links
- AgentGid page: https://agentgid.com/assets/fastvideo/
- Is it free for commercial use? https://agentgid.com/assets/fastvideo/commercial-use/
- Source: https://huggingface.co/FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree
- Source: https://huggingface.co/FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree/blob/main/LICENSE


## FastH3 Preview: fast open model for video with sound

FastH3 Preview is an open AI model from the FastVideo team that makes a short video with matching sound, such as voices and noises, from a written description. It is a sped-up version of MiniMax's H3 video model and needs very powerful data-center graphics cards. Its license does not allow any use in the European Union, the United Kingdom, South Korea or the United States.

### Where to use it
- A clip with sound from text: Describe a scene, the camera and the sounds, and get one video where the picture, speech and noises fit together.
- Vertical and square videos: Make the same kind of clip upright for phone stories, square for a feed, or wide for a website, by setting the size.
- Run on other multi-card servers: On a rented server with several NVIDIA cards that are not B200, switch the fast attention parts to the general versions.
- Download only the small adapter: If you already have the base H3 model, take only the small speed-up add-on instead of all the files again.
- Train your own faster video model: Teams that build video models can use the same FastVideo toolkit to fine-tune and speed up other open video models.
- Explainer clips with narration: Make short how-to or product clips where the narrator's voice and the on-screen action match, for teams inside the allowed countries. (idea, not from the official pages)

### How to start
- Get the FastVideo code and install it with its FastH3 extras in a fresh Python 3.12 environment (path for CUDA 13 and Blackwell cards).

```
git clone https://github.com/hao-ai-lab/FastVideo.git
cd FastVideo
uv venv --python 3.12 --seed
source .venv/bin/activate
UV_TORCH_BACKEND=cu130 uv pip install \
  --no-sources-package fastvideo-kernel \
  -e ".[fasth3]"
```

- Run the example script with this model and your prompt; the files download on first run.

```
python examples/inference/basic/basic_fasth3.py --model-path FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree --prompt "your prompt" --no-warmup --repeats 1
```


### License
Under the MiniMax H3 Community License you may use it, also commercially, only outside the EU, UK, South Korea and the USA. Paid products must show "MiniMax H3" in their interface, and outputs may not train other AI models.

### Good to know
- The files are huge, around 148 GB in total, with the main model and the text reader each split into fourteen parts of about 5 GB.
- This is a preview. Hard motion, fine detail and some sound may be weaker than the full H3 model, and starting from an image is not supported.
- Companies with more than 20 million US dollars of yearly revenue need written permission from MiniMax first.
- If you let other people make videos with it, you must run safety checks, take reports of misuse and bind users to the same rules.
- Videos posted publicly must be clearly marked as machine-made.

Checked 2026-11-12 against: https://huggingface.co/FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree, https://huggingface.co/FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree/raw/main/README.md, https://huggingface.co/FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree/raw/main/LICENSE, https://huggingface.co/api/models/FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree, https://huggingface.co/api/models/FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree/tree/main?recursive=true, https://github.com/hao-ai-lab/FastVideo, https://github.com/hao-ai-lab/FastVideo/blob/main/README.md, https://github.com/hao-ai-lab/FastVideo/blob/main/examples/inference/basic/basic_fasth3.py, https://github.com/hao-ai-lab/FastVideo/blob/main/LICENSE, https://haoailab.com/blogs/fasth3-preview/

### Mistakes to avoid
- Don't: running basic_fasth3.py without --model-path → Do: always pass --model-path FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree (The script's default points to a different, older preview checkpoint.)
- Don't: swapping in a standard attention backend to avoid installing the VSA kernel → Do: install FastVideo's VSA-H3 backend and kernel (The model needs that backend for the stated speed and quality.)
- Don't: --num-gpus 3, 5 or 6 → Do: use a card count that divides 56, such as 2, 4, 7 or 8 (The model's work is split across cards by its 56 attention heads.)
- Don't: --height 720 --width 1280 → Do: sizes in multiples of 32, such as 768 by 1344 (Other sizes are not accepted.)
- Don't: setting it up for a person or company in the EU, UK, South Korea or USA → Do: stop and pick a model without territory limits, such as Wan 2.2 (The license does not allow any use there, outputs included.)

### Questions
- Can I use it from Germany, the UK or the USA? No. The license does not apply in the EU, the UK, South Korea or the USA; MiniMax invites people there to ask about a separate license.
- Who owns the videos I make? MiniMax claims no rights over them; you and your users are responsible for them and their use.
- Can it start from my own picture? Not this preview; it supports text to video with sound only.
- Does it run on one home graphics card? The tested setup is four B200 data-center cards; on other multi-card systems the card count must divide 56.

---
AgentGid's catalog is curated: every entry was checked against the maker's official pages (registry, repository, license and pricing pages) on the date shown; a license value shows the quote it stands on wherever AgentGid has one. Usage numbers come from AgentGid's own scan of public repositories set up for AI coding agents. Claims by makers are reported, not tested by AgentGid.
